3 # Any copyright is dedicated to the Public Domain.
4 # https://creativecommons.org/publicdomain/zero/1.0/
6 # Written by Francois Fleuret <francois@fleuret.org>
8 import math, sys, argparse, time, tqdm, os, datetime, warnings
10 import torch, torchvision
12 from torch.nn import functional as F
15 import mygpt, tasks, problems
17 ######################################################################
19 if torch.cuda.is_available():
20 device = torch.device("cuda")
21 torch.backends.cuda.matmul.allow_tf32 = True
23 device = torch.device("cpu")
25 ######################################################################
27 parser = argparse.ArgumentParser(
28 description="An implementation of GPT with cache.",
29 formatter_class=argparse.ArgumentDefaultsHelpFormatter,
36 help="file, byheart, learnop, guessop, mixing, memory, twotargets, addition, picoclvr, mnist, maze, snake, stack, expr, rpl, grid, qmlp, greed",
39 parser.add_argument("--log_filename", type=str, default="train.log", help=" ")
41 parser.add_argument("--result_dir", type=str, default=None)
43 parser.add_argument("--seed", type=int, default=0)
45 parser.add_argument("--max_percents_of_test_in_train", type=int, default=1)
47 ########################################
49 parser.add_argument("--nb_epochs", type=int, default=50)
51 parser.add_argument("--batch_size", type=int, default=None)
53 parser.add_argument("--physical_batch_size", type=int, default=None)
55 parser.add_argument("--nb_train_samples", type=int, default=None)
57 parser.add_argument("--nb_test_samples", type=int, default=None)
59 parser.add_argument("--optim", type=str, default="adam")
61 parser.add_argument("--learning_rate", type=float, default=1e-4)
63 parser.add_argument("--learning_rate_schedule", type=str, default="10: 2e-5,30: 4e-6")
65 ########################################
67 parser.add_argument("--model", type=str, default=None)
69 parser.add_argument("--dim_model", type=int, default=None)
71 parser.add_argument("--dim_keys", type=int, default=None)
73 parser.add_argument("--dim_hidden", type=int, default=None)
75 parser.add_argument("--nb_heads", type=int, default=None)
77 parser.add_argument("--nb_blocks", type=int, default=None)
79 parser.add_argument("--dropout", type=float, default=0.1)
81 ########################################
83 parser.add_argument("--deterministic_synthesis", action="store_true", default=False)
85 ##############################
88 parser.add_argument("--filetask_train_file", type=str, default=None)
90 parser.add_argument("--filetask_test_file", type=str, default=None)
92 ##############################
95 parser.add_argument("--rpl_nb_starting_values", type=int, default=3)
97 parser.add_argument("--rpl_max_input", type=int, default=9)
99 parser.add_argument("--rpl_prog_len", type=int, default=8)
101 parser.add_argument("--rpl_nb_runs", type=int, default=5)
103 parser.add_argument("--rpl_no_prog", action="store_true", default=False)
105 ##############################
108 parser.add_argument("--grid_size", type=int, default=6)
110 parser.add_argument("--grid_fraction_play", type=float, default=0)
112 ##############################
115 parser.add_argument("--picoclvr_nb_colors", type=int, default=5)
117 parser.add_argument("--picoclvr_height", type=int, default=12)
119 parser.add_argument("--picoclvr_width", type=int, default=16)
121 parser.add_argument("--picocvlr_prune_properties", type=str, default="none")
123 ##############################
126 parser.add_argument("--maze_height", type=int, default=13)
128 parser.add_argument("--maze_width", type=int, default=21)
130 parser.add_argument("--maze_nb_walls", type=int, default=15)
132 ##############################
135 parser.add_argument("--snake_height", type=int, default=9)
137 parser.add_argument("--snake_width", type=int, default=12)
139 parser.add_argument("--snake_nb_colors", type=int, default=5)
141 parser.add_argument("--snake_length", type=int, default=200)
143 ##############################
146 parser.add_argument("--byheart_separation", type=int, default=1)
148 ##############################
151 parser.add_argument("--stack_nb_steps", type=int, default=100)
153 parser.add_argument("--stack_nb_stacks", type=int, default=3)
155 parser.add_argument("--stack_nb_digits", type=int, default=3)
157 parser.add_argument("--stack_fraction_values_for_train", type=float, default=None)
159 ##############################
162 parser.add_argument("--expr_nb_variables", type=int, default=5)
164 parser.add_argument("--expr_sequence_length", type=int, default=40)
166 parser.add_argument("--expr_operand_max", type=int, default=9)
168 parser.add_argument("--expr_result_max", type=int, default=99)
170 parser.add_argument("--expr_input_file", type=str, default=None)
172 ##############################
175 parser.add_argument("--mixing_hard", action="store_true", default=False)
177 parser.add_argument("--mixing_deterministic_start", action="store_true", default=False)
179 ##############################
182 parser.add_argument("--greed_height", type=int, default=5)
184 parser.add_argument("--greed_width", type=int, default=7)
186 parser.add_argument("--greed_T", type=int, default=25)
188 parser.add_argument("--greed_nb_walls", type=int, default=5)
190 parser.add_argument("--greed_nb_coins", type=int, default=2)
192 ######################################################################
194 args = parser.parse_args()
196 assert args.picocvlr_prune_properties in {"none", "train+eval", "eval"}
198 if args.result_dir is None:
199 args.result_dir = f"results_{args.task}"
201 ######################################################################
203 default_task_args = {
207 "nb_train_samples": 250000,
208 "nb_test_samples": 10000,
213 "nb_train_samples": 250000,
214 "nb_test_samples": 10000,
219 "nb_train_samples": 250000,
220 "nb_test_samples": 10000,
225 "nb_train_samples": 50000,
226 "nb_test_samples": 10000,
231 "nb_train_samples": 2500000,
232 "nb_test_samples": 10000,
237 "nb_train_samples": 250000,
238 "nb_test_samples": 10000,
243 "nb_train_samples": 100000,
244 "nb_test_samples": 1000,
249 "nb_train_samples": 1000000,
250 "nb_test_samples": 10000,
255 "nb_train_samples": 50000,
256 "nb_test_samples": 10000,
261 "nb_train_samples": 100000,
262 "nb_test_samples": 10000,
267 "nb_train_samples": 250000,
268 "nb_test_samples": 10000,
273 "nb_train_samples": 2500000,
274 "nb_test_samples": 10000,
279 "nb_train_samples": 250000,
280 "nb_test_samples": 10000,
285 "nb_train_samples": 100000,
286 "nb_test_samples": 1000,
291 "nb_train_samples": 50000,
292 "nb_test_samples": 10000,
297 "nb_train_samples": 25000,
298 "nb_test_samples": 1000,
303 "nb_train_samples": 250000,
304 "nb_test_samples": 10000,
309 "nb_train_samples": 60000,
310 "nb_test_samples": 10000,
315 "nb_train_samples": 25000,
316 "nb_test_samples": 10000,
320 if args.task in default_task_args:
321 for k, v in default_task_args[args.task].items():
322 if getattr(args, k) is None:
325 ######################################################################
327 default_model_args = {
365 if args.model in default_model_args:
366 for k, v in default_model_args[args.model].items():
367 if getattr(args, k) is None:
370 raise ValueError(f"Unknown model {args.model}")
372 ######################################################################
375 os.mkdir(args.result_dir)
376 except FileExistsError:
378 print(f"result directory {args.result_dir} already exists")
381 log_file = open(os.path.join(args.result_dir, args.log_filename), "a")
384 # torch.backends.cudnn.deterministic = True
385 # torch.backends.cudnn.benchmark = False
386 # torch.use_deterministic_algorithms(True)
387 torch.manual_seed(args.seed)
388 if torch.cuda.is_available():
389 torch.cuda.manual_seed_all(args.seed)
391 ######################################################################
395 t = time.strftime("%Y%m%d-%H:%M:%S ", time.localtime())
397 if log_file is not None:
398 log_file.write(t + s + "\n")
405 log_string(f"argv {' '.join(sys.argv)}")
408 log_string(f"args.{n} {getattr(args, n)}")
411 ######################################################################
414 def picoclvr_pruner_horizontal_green(p):
415 return not ("green" in p and ("left" in p or "right" in p))
418 picoclvr_pruner_train = (
419 picoclvr_pruner_horizontal_green
420 if args.picocvlr_prune_properties in {"train+eval"}
424 picoclvr_pruner_eval = (
425 (lambda p: not picoclvr_pruner_horizontal_green(p))
426 if args.picocvlr_prune_properties in {"train+eval", "eval"}
430 ######################################################################
432 if args.physical_batch_size is None:
433 args.physical_batch_size = args.batch_size
435 assert args.batch_size % args.physical_batch_size == 0
437 assert args.nb_train_samples % args.batch_size == 0
438 assert args.nb_test_samples % args.batch_size == 0
440 if args.task == "file":
442 args.filetask_train_file is not None and args.filetask_test_file is not None
443 ), "You have to specify the task train and test files"
444 task = tasks.TaskFromFile(
445 args.filetask_train_file,
446 args.filetask_test_file,
447 nb_train_samples=args.nb_train_samples,
448 nb_test_samples=args.nb_test_samples,
449 batch_size=args.physical_batch_size,
453 args.max_percents_of_test_in_train = 0
455 elif args.task == "byheart":
456 task = tasks.SandBox(
457 problem=problems.ProblemByHeart(separation=args.byheart_separation),
458 nb_train_samples=args.nb_train_samples,
459 nb_test_samples=args.nb_test_samples,
460 batch_size=args.physical_batch_size,
464 args.max_percents_of_test_in_train = -1
466 elif args.task == "world":
468 nb_train_samples=args.nb_train_samples,
469 nb_test_samples=args.nb_test_samples,
470 batch_size=args.physical_batch_size,
471 result_dir=args.result_dir,
475 args.max_percents_of_test_in_train = -1
477 elif args.task == "learnop":
478 task = tasks.SandBox(
479 problem=problems.ProblemLearnOperator(),
480 nb_train_samples=args.nb_train_samples,
481 nb_test_samples=args.nb_test_samples,
482 batch_size=args.physical_batch_size,
488 elif args.task == "guessop":
489 task = tasks.SandBox(
490 problem=problems.ProblemGuessOperator(),
491 nb_train_samples=args.nb_train_samples,
492 nb_test_samples=args.nb_test_samples,
493 batch_size=args.physical_batch_size,
499 elif args.task == "twotargets":
500 task = tasks.SandBox(
501 problem=problems.ProblemTwoTargets(),
502 nb_train_samples=args.nb_train_samples,
503 nb_test_samples=args.nb_test_samples,
504 batch_size=args.physical_batch_size,
509 elif args.task == "memory":
510 task = tasks.SandBox(
511 problem=problems.ProblemMemory(),
512 nb_train_samples=args.nb_train_samples,
513 nb_test_samples=args.nb_test_samples,
514 batch_size=args.physical_batch_size,
519 elif args.task == "mixing":
520 task = tasks.SandBox(
521 problem=problems.ProblemMixing(
522 hard=args.mixing_hard, random_start=not args.mixing_deterministic_start
524 nb_train_samples=args.nb_train_samples,
525 nb_test_samples=args.nb_test_samples,
526 batch_size=args.physical_batch_size,
531 elif args.task == "addition":
532 task = tasks.SandBox(
533 problem=problems.ProblemAddition(),
534 nb_train_samples=args.nb_train_samples,
535 nb_test_samples=args.nb_test_samples,
536 batch_size=args.physical_batch_size,
541 elif args.task == "picoclvr":
542 task = tasks.PicoCLVR(
543 nb_train_samples=args.nb_train_samples,
544 nb_test_samples=args.nb_test_samples,
545 batch_size=args.physical_batch_size,
546 height=args.picoclvr_height,
547 width=args.picoclvr_width,
548 nb_colors=args.picoclvr_nb_colors,
551 pruner_train=picoclvr_pruner_train,
552 pruner_eval=picoclvr_pruner_eval,
555 elif args.task == "mnist":
557 nb_train_samples=args.nb_train_samples,
558 nb_test_samples=args.nb_test_samples,
559 batch_size=args.physical_batch_size,
563 elif args.task == "maze":
565 nb_train_samples=args.nb_train_samples,
566 nb_test_samples=args.nb_test_samples,
567 batch_size=args.physical_batch_size,
568 height=args.maze_height,
569 width=args.maze_width,
570 nb_walls=args.maze_nb_walls,
574 elif args.task == "snake":
576 nb_train_samples=args.nb_train_samples,
577 nb_test_samples=args.nb_test_samples,
578 batch_size=args.physical_batch_size,
579 height=args.snake_height,
580 width=args.snake_width,
581 nb_colors=args.snake_nb_colors,
582 length=args.snake_length,
583 prompt_length=args.snake_length // 2,
587 elif args.task == "stack":
589 nb_train_samples=args.nb_train_samples,
590 nb_test_samples=args.nb_test_samples,
591 batch_size=args.physical_batch_size,
593 nb_steps=args.stack_nb_steps,
594 nb_stacks=args.stack_nb_stacks,
595 nb_digits=args.stack_nb_digits,
596 fraction_values_for_train=args.stack_fraction_values_for_train,
600 elif args.task == "expr":
602 nb_train_samples=args.nb_train_samples,
603 nb_test_samples=args.nb_test_samples,
604 nb_variables=args.expr_nb_variables,
605 sequence_length=args.expr_sequence_length,
606 operand_max=args.expr_operand_max,
607 result_max=args.expr_result_max,
608 batch_size=args.physical_batch_size,
612 elif args.task == "rpl":
614 nb_train_samples=args.nb_train_samples,
615 nb_test_samples=args.nb_test_samples,
616 batch_size=args.physical_batch_size,
617 nb_starting_values=args.rpl_nb_starting_values,
618 max_input=args.rpl_max_input,
619 prog_len=args.rpl_prog_len,
620 nb_runs=args.rpl_nb_runs,
621 no_prog=args.rpl_no_prog,
626 elif args.task == "grid":
628 nb_train_samples=args.nb_train_samples,
629 nb_test_samples=args.nb_test_samples,
630 batch_size=args.physical_batch_size,
632 fraction_play=args.grid_fraction_play,
637 elif args.task == "qmlp":
639 nb_train_samples=args.nb_train_samples,
640 nb_test_samples=args.nb_test_samples,
641 batch_size=args.physical_batch_size,
642 result_dir=args.result_dir,
647 elif args.task == "greed":
649 nb_train_samples=args.nb_train_samples,
650 nb_test_samples=args.nb_test_samples,
651 batch_size=args.physical_batch_size,
652 height=args.greed_height,
653 width=args.greed_width,
655 nb_walls=args.greed_nb_walls,
656 nb_coins=args.greed_nb_coins,
662 raise ValueError(f"Unknown task {args.task}")
664 ######################################################################
666 log_string(f"device {device}")
668 vocabulary_size = task.vocabulary_size()
670 log_string(f"vocabulary_size {vocabulary_size}")
672 ##############################
679 vocabulary_size=vocabulary_size,
680 dim_model=args.dim_model,
681 dim_keys=args.dim_keys,
682 dim_hidden=args.dim_hidden,
683 nb_heads=args.nb_heads,
684 nb_blocks=args.nb_blocks,
686 dropout=args.dropout,
691 nb_parameters = sum(p.numel() for p in models[0].parameters())
692 log_string(f"nb_parameters {nb_parameters} ({int(nb_parameters/1e6)}M)")
694 ######################################################################
696 # Compute the entropy of the training tokens
699 for input in task.batches(split="train", desc="train-entropy"):
700 token_count += F.one_hot(input, num_classes=task.vocabulary_size()).sum((0, 1))
701 token_probas = token_count / token_count.sum()
702 entropy = -torch.xlogy(token_probas, token_probas).sum()
703 train_set_perplexity = math.exp(entropy)
705 ######################################################################
706 # A bit of paranoia never hurts
708 if args.max_percents_of_test_in_train >= 0:
710 def subsets_as_tuples(batches, cs):
712 for batch in batches:
714 s.add(tuple([v.item() for v in x]))
720 nb_test, nb_in_train = 0, 0
721 for test_subset in subsets_as_tuples(
722 task.batches(split="test", desc="test-check"), 25000
725 for train_subset in subsets_as_tuples(
726 task.batches(split="train", desc="train-check"), 25000
728 in_train.update(test_subset.intersection(train_subset))
729 nb_in_train += len(in_train)
730 nb_test += len(test_subset)
733 f"data_check {nb_in_train*100/nb_test:.02f}% ({nb_in_train}/{nb_test}) of test samples are in the train set"
737 nb_in_train <= args.max_percents_of_test_in_train * nb_test / 100
738 ), f"More than {args.max_percents_of_test_in_train}% of test samples are in the train set"
740 ##############################
742 if args.learning_rate_schedule == "cos":
743 learning_rate_schedule = {}
744 for n_epoch in range(args.nb_epochs):
745 u = n_epoch / args.nb_epochs * math.pi
746 learning_rate_schedule[n_epoch] = args.learning_rate * 0.5 * (1 + math.cos(u))
751 tuple(x.split(":")) for x in args.learning_rate_schedule.split(",")
755 learning_rate_schedule = {}
756 learning_rate = args.learning_rate
757 for n_epoch in range(args.nb_epochs):
759 learning_rate = u[n_epoch]
760 learning_rate_schedule[n_epoch] = learning_rate
762 log_string(f"learning_rate_schedule {learning_rate_schedule}")
764 time_pred_result = None
766 ######################################################################
769 def one_epoch(model, task, learning_rate):
770 log_string(f"learning_rate {learning_rate}")
772 if args.optim == "sgd":
773 optimizer = torch.optim.SGD(model.parameters(), lr=learning_rate)
774 elif args.optim == "adam":
775 optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
776 elif args.optim == "adamw":
777 optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
779 raise ValueError(f"Unknown optimizer {args.optim}.")
783 nb_train_samples, acc_train_loss = 0, 0.0
785 for input in task.batches(split="train"):
786 input = input.to(device)
788 if nb_train_samples % args.batch_size == 0:
789 optimizer.zero_grad()
791 output = model(mygpt.BracketedSequence(input)).x
792 loss = F.cross_entropy(output.transpose(1, 2), input)
793 acc_train_loss += loss.item() * input.size(0)
795 nb_train_samples += input.size(0)
799 if nb_train_samples % args.batch_size == 0:
802 train_perplexity = math.exp(min(100, acc_train_loss / nb_train_samples))
804 log_string(f"train_perplexity {n_epoch} {train_perplexity}")
807 ######################################################################
810 def run_tests(model, task, deterministic_synthesis):
811 with torch.autograd.no_grad():
814 nb_test_samples, acc_test_loss = 0, 0.0
815 nb_samples_accumulated = 0
817 for input in task.batches(split="test"):
818 input = input.to(device)
820 bs = model(mygpt.BracketedSequence(input))
823 loss = F.cross_entropy(output.transpose(1, 2), input)
825 acc_test_loss += loss.item() * input.size(0)
827 nb_test_samples += input.size(0)
829 main_test_accuracy = task.produce_results(
832 result_dir=args.result_dir,
834 deterministic_synthesis=deterministic_synthesis,
837 test_perplexity = math.exp(min(100, acc_test_loss / nb_test_samples))
839 log_string(f"test_perplexity {n_epoch} {test_perplexity}")
841 return main_test_accuracy
844 ######################################################################
859 while sum([x.size(0) for x in kept]) < nb_for_train + nb_for_test:
860 new_quizzes, nb_correct = task.create_new_quizzes(
862 result_dir=args.result_dir,
864 nb=4 * (nb_for_train + nb_for_test),
866 other_models=other_models,
870 to_keep = new_quizzes[
872 nb_correct >= nb_min_correct, nb_correct <= nb_max_correct
875 log_string(f"keep {to_keep.size(0)} quizzes")
878 new_quizzes = torch.cat(kept, dim=0)[: nb_for_train + nb_for_test]
880 task.store_new_quizzes(new_quizzes[:nb_for_train], for_train=True)
881 task.store_new_quizzes(new_quizzes[nb_for_train:], for_train=False)
886 f"world_new_{n_epoch:04d}.png",
891 ######################################################################
893 accuracy_to_make_quizzes = 0.95
895 for n_epoch in range(nb_epochs_finished, args.nb_epochs):
896 learning_rate = learning_rate_schedule[n_epoch]
899 one_epoch(m, task, learning_rate)
900 test_accuracy = run_tests(m, task, deterministic_synthesis=False)
902 if test_accuracy >= accuracy_to_make_quizzes:
903 other_models = models.copy()
904 other_models.remove(model)
905 create_quizzes(other_models, task)
907 # --------------------------------------------
909 time_current_result = datetime.datetime.now()
910 if time_pred_result is not None:
912 f"next_result {time_current_result + (time_current_result - time_pred_result)}"
914 time_pred_result = time_current_result
916 ######################################################################